Most writing about AI in SEO falls into two camps. Either it is breathless — "AI will replace SEO teams" — or defensive — "AI can't understand strategy." Both are avoiding the interesting question, which is: what specifically changed in the daily work?
I have been running Claude Code as part of my daily workflow for about a year, on a site with roughly 150,000 daily organic users across 900+ locations. Here is the honest accounting.
What it replaced
It replaced the scripted middle layer — the work that was too repetitive to be interesting and too bespoke to justify buying a tool for. Log parsing, schema auditing across large page sets, sitemap diffing, crawl export reshaping, and validation scripting. Tasks with verifiable outputs, where I can check the answer against source data.
The pattern that predicts success is verifiability. If I can look at the output and confirm it against the raw data in under a minute, agentic execution works. If verifying the output requires the same effort as producing it, the automation is theater.
| Task | Before | Now | Verdict |
|---|---|---|---|
| AI crawler log analysis GPTBot, ClaudeBot, PerplexityBot hit patterns by section |
~4 hrs/month in spreadsheets | ~20 min, scripted | Replaced |
| Schema audit across templates Which page types have which types, and what's malformed |
Manual spot checks, incomplete | Full coverage, scripted | Replaced + improved |
| Sitemap diffing between crawls | ~1 hr, error-prone | Minutes | Replaced |
| Crawl export reshaping Botify/Screaming Frog exports into reporting format |
Tedious, every time | One-time script, reused | Replaced |
| Prompt-set citation testing | ~2 hrs/month manual | Still ~2 hrs — deliberately | Kept manual |
| Deciding what to fix first | Judgment | Judgment | Not automatable |
| Stakeholder narrative Why this matters, in dollars, to a VP |
Judgment | Judgment | Not automatable |
The one I deliberately did not automate
Prompt-set citation testing is the obvious automation candidate. Run 25 prompts across four engines, parse the responses, score citations, write to a sheet. It is mechanical. I could have scripted it in an afternoon.
I still do it by hand, and I will keep doing it by hand.
The reason is that reading the actual generated answers teaches you things the score does not. Last quarter I noticed that when our brand did get cited in a particular category, the framing consistently positioned us as the budget option — accurate for one product line, wrong for the business. A citation-rate number would have shown that quarter as a win. Reading the answers showed a positioning problem that was invisible in the metric.
Automate the work where the output is the answer. Keep the work manual where the output is a number and the answer is in the texture underneath it.
Where I got burned
Early on I had an agent generate a set of schema blocks for a large template group. The output was clean, well-formed, and validated against the Rich Results Test. It was also subtly wrong — it had inferred product attributes from page copy rather than pulling from the actual product feed, so a meaningful subset carried inaccurate structured data.
Nothing about the output looked wrong. It validated. It parsed. It was plausible. That is precisely the failure mode: the risk is not bad output, it is plausible output approved without verification against source data.
What I do now: any generated artifact touching production goes through the same review gate a junior engineer's pull request would. Spot-check against source, not against format validity. Format validity is what the agent optimizes for, so format validity is exactly the wrong thing to verify on.
A workflow worth stealing
The single highest-value thing I built is a monthly AI crawler report. It answers one question: where are AI crawlers spending their budget on our site, and is that where we want them?
1. Filter server logs to GPTBot, ClaudeBot, PerplexityBot, Google-Extended
2. Group hits by URL section (/product/, /blog/, /resources/, /solutions/)
3. Join against revenue or conversion value per section
4. Flag: high-value sections with low AI crawl share
5. Flag: high AI crawl share on low-value sections
6. Diff against prior month — what shifted, and why
Step 4 is the one that produces action. The first time I ran it, AI crawlers were spending the large majority of their budget on our resource and blog sections and barely touching product and solution pages — the inverse of what we wanted for citation pickup on commercial queries. That is not a finding a rank tracker surfaces. It came out of logs nobody was reading because reading them was tedious.
Tedium was the only thing protecting that insight. Removing the tedium surfaced it.
The thing that did not change
Agentic tooling compressed execution time substantially. It did not change what was worth executing. Every month the meeting where we decide what to work on runs about as long as it used to, and it is still the meeting that determines whether the quarter works.
The value of an SEO practitioner was never typing speed. It was knowing which of forty possible fixes will move revenue and being able to explain that to someone who controls budget. Those remain scarce, and if anything they got scarcer — because when execution is cheap, the cost of executing the wrong thing goes up, not down.
One caution on content
Using AI tooling in your workflow is not the same as publishing AI-generated content at volume. The 2026 core update data showed measurable penalties for thin content produced at scale without human editorial judgment applied to the output — not for AI-assisted content generally, but for content where nobody with domain expertise reviewed or owned the result.
Use the tooling for analysis, transformation, and scripting. Keep a human accountable for anything that publishes. I cover the citation side of that distinction in more depth over at Sable Search.
Curious where your brand stands in AI search?
Free 5-prompt snapshot through Sable Search — tested across ChatGPT, Claude, Perplexity, and Google AI Overviews against your top competitor. One page, 48 hours.
Get the free snapshot ↗Frequently asked questions
Claude Code handles the scripted, repeatable layer well — parsing server logs for AI crawler behavior, auditing schema across large page sets, diffing sitemaps between crawls, reformatting crawl exports, and generating validation scripts. It works because these tasks have verifiable outputs. It does not replace judgment about what to prioritize or why a pattern matters.
It is safe for analysis and generation, and unsafe for unreviewed deployment. Every output that touches a live site should pass through the same review process a junior engineer's pull request would. The failure mode is not the agent producing bad work — it is a human approving plausible-looking output without verifying it against the actual source data.
Using AI tools in your workflow does not hurt visibility. Publishing unreviewed AI-generated content at scale does. The 2026 core update data showed measurable penalties for content produced at volume without human editorial judgment applied to the output. The distinction that matters is whether a human with domain expertise reviewed and owns the result.